scaling, renormalization and self- similarity in complex networks chaoming song (ccny) lazaros...

Scaling, renormalization and self-similarity in complex networks

Chaoming Song (CCNY)Chaoming Song (CCNY)Lazaros Gallos (CCNY)Lazaros Gallos (CCNY)Shlomo Havlin (Bar-Ilan, Israel)Shlomo Havlin (Bar-Ilan, Israel)

Hernan A. MakseHernan A. Makse

Levich Institute and Physics Dept.Levich Institute and Physics Dept.City College of New YorkCity College of New York

Protein interaction networkProtein interaction network

Are “scale-free” networks really ‘free-of-scale’?“If you had asked me yesterday, I would have said surely not” - said Barabasi.

(Science News, February 2, 2005).

Small world contradicts self-similarity!

Small World effect shows that distance between nodes grows logarithmically with N (the network size):

OR

Self-similar = fractal topology is defined by a power-law relation:

AIM: How the network behaves under a scale transformation.Implications for: 1. Dynamics 2. Modularity

3. Universality

WWW nd.edu

300,000 web-pages

Internet connectivity, with selected backbone ISPs (Internet Service Provider) colored separately.

Faloutsos et al., SIGCOMM ’99

Internet

J. Han et al., Nature (2004)

Yeast Protein-Protein Interaction Map

Individual proteins

Physical interactions from the “filtered yeast interactome” database: 2493 high-confidence interactions observed by at least two methods (yeast two-hybrid).1379 proteins, <k> = 3.6

Colored according to protein function in the cell:Transcription, Translation, Transcription control, Protein-fate, Genome maintenance, Metabolism, Unknown, etc

Modular structure according to function!

from MIPS database, mips.gsf.de

Metabolic network of biochemical reactions in E.coli

Chemical substrates

Biochemical interactions: enzyme-catalyzed reactions that transform one metabolite into another.

Modular structureaccording to the biochemicalclass of the metabolic productsof the organism.

Colored according to product class:Lipids, essential elements, protein, peptides and amino acids, coenzymes and prosthetic groups, carbohydrates, nucleotides and nucleic acids.

J. Jeong, et al., Nature, 407 651 (2000)

Biological networks

Protein Homology Tree of life

Similarities between sequence ofAmino-acids (BLAST)Network of 5 million proteins1.2 TB of data growing at 50GBPer month.Adai et al. J Mol Biol (2004)

Complex network of speciesRepresenting their evolucionary history~90,000 species

Coast lines Rivers Mountains

Clouds Lightening Neurons

Introduction to fractalsIn Nature there exist many examples of random fractals

How long is the coastline of Norway?It depends on the length of your ruler.

Fractal Dimension dB-Box Covering Method

Fractals look the same on all scales = `scale-invariant’.

Box length

Total no. of boxes

Boxing in Biology

How to “zoom out” of a complex network?

Generate boxes where all nodes are within a distance

Calculate number of boxes, , of size needed to cover the network

We need the minimum number of boxes: NP-complete optimization problem!We need the minimum number of boxes: NP-complete optimization problem!

Boxing in Biology

Most efficient tiling of the network

4 boxes

5 boxes

1

0

0

1

2

8 node network: Easy to solve

300,000 node network: Mapping to graph colouring problem. NP-complete: Greedy algorithm to find minimum boxes

Burning algorithms

1. Compact box burning: CBBSong et al. JSTAT (2007)

2. Maximum mass burning: MEMB Burning from the hubs with the radius r

Minimazing the number of boxes is analogous to maximizing the mass of each box: implications for modularity

Two universality classes:

-dB

log(lB)

log(

NB)

1 2 3

TOPOLOGICAL NON-FRACTAL TOPOLOGICAL FRACTALS

EUCLIDIAN NON-FRACTALS EUCLIDIAN FRACTALS

Percolation cluster:“holes” at all scales

Compact cluster

Box covering in yeast: protein interaction network

Many complex networks are Fractal

Metabolic Protein interaction

Song, Havlin, Makse, Nature (2005)

Biological networks

Three domains of life: archaea, bacteria, eukaria

E. coli, H. sapiens, yeast

43 organisms - all scale

yeast

Metabolic networks are fractals

More topological fractals

WWW

nd.edu domain

1. Protein homology network2. Tree of life (taxonomy)3. Genetic networks (Meyer-Ortmanns, Khang)4. Neural networks (Yuste)

300,000 web-pages

Internet and social networks are not fractal

Other models fail too: Erdos-Renyi, hierarchical model, fitness model, JKK model, pseudo-fractals models, etc.

The Barabasi-Albert model of preferentialattachment does not generate fractal networks

All available models fail to predict self-similarity

INTERNETRouter and AS level

Two universality classes

Fractal networks:WWWBiological networks: protein interactions, metabolic, genetic (Meyer-Ortmanns, Khang), taxonomy, tree of life, protein homology network, neural activity network.

Non-Fractal networks: Internet (routers and AS level)Social networks (citations (Khang), IMDB)Models based on uncorrelated preferential attachment

Two ways to calculate fractal dimensions

Box covering method Cluster growing method

In homogeneous systems (all nodes with similar k) both definitions agree:

percolation

Box Covering= flat average Cluster Growing = biased

power law

Different methods yield different results due to heterogeneous topology

exponential

Box covering reveals the self similarity. Cluster growth reveals the small world. NO CONTRADICTION! SAME HUBS ARE USED MANY TIMES IN CG.

Renormalization in Complex Networks

NOW, REGARD EACH BOX AS A SINGLE NODEAND ASK WHAT IS THE DEGREEDISRIBUTION OF THE NETWORKOF BOXES AT DIFFERENT SCALES ?

Renormalization of WWW network with

Statistical properties are invariant under renormalization

WWW PIN

E.coli

Internet

Self-similarity:Invariant under renormalization

Internet is not fractal, dB--> infinity but it is renormalizable

FRACTALS NON-FRACTALS

DYNAMICS: Turning back the timeRepeatedly BOXING the network is the same as going back

in time: from a single node to present day.

renormalization

time evolution

Can we “predict” the past…. ? if not the future.

ancestral node

present daynetwork

THE RENORMALIZATION SCHEME

1

time evolution

Evolution of complex networks

opening boxes

How does Modularity arise?The boxes have a physical meaning =

self-similar nested communities

time evolution

ancestral node

present daynetwork

renormalization

1

How to identify communities in complex networks?

Classes of genes in the yeast proteome

Is evolution of PIN fractal?

Ancestral Prokaryote Cell

YeastOtherFungi

Ancestral yeast

Animals+ Plants

Ancestral Fungus

Archaea + Bacteria

Ancestral Eukaryote

presentday

~ 300 million years ago

1 billion years ago

1.5 billion years ago

Following the phylogenetic tree of life:

3.5 billion years ago

COG databasevon Mering, et al Nature (2002)

Suggests that present-day networks could have been created following a self-similar, fractal dynamics.

Same fractal dimension and scale-freeexponent over 3.5 billion years…

Renormalization following the phylogenetic treeRenormalization following the phylogenetic tree

P. Uetz, et al. Nature 403 (2000).

Emergence of Modularity in PINBoxes are related to the biologically relevant functional modules

in the yeast protein interactome

time evolution renormalization

present day network

translation transcription protein-fatecellular-fateorganization

ancestralcell

Emergence of modularity in metabolic networks

Appearance of functional modules in E. coli metabolic network.Most robust network than non-fractals.

Scaling theory of modularityHow the modules/communities are linked?

k: degree of the nodes

k’=2renormalization

s=1/4k=8

k’: degree of the communities

node degree

community degree factor<1

Gallos et al. PNAS (2007)

Theoretical approach to modular networks: Scaling theory to the rescue

WWW

The larger the modulethe smaller their connectivity

new exponent describing how modules link

Scaling relations

A theoretical prediction relating the different exponents

new scaling relation

boxes

distance

degree

new exponent

Scaling relationsThe communities also follow a self-similar pattern

WWW Metabolic

Scaling relationworks

fractalsfractals communities/modulescommunities/modules

scale-freescale-free

predictionprediction

What is the origin of topological fractality?

HINT: the key to understand fractals is in the degreecorrelations P(k1,k2) not in P(k)

Can you see the difference?

Internet map Yeast protein map

E.coli metabolic map

NON FRACTAL FRACTAL

Compact cluster

Quantifying correlations P(k1,k2):

Probability to find a node with k1 links connected with a node of k2 links

Internet map - non fractal Metabolic map - fractal

log(k1)log(k1)

log(

k 2)

log(

k 2)

P(k1,k2)

low prob.

low prob.

high prob.

high prob.

Hubs connected with hubs Hubs connected with non-hubs

Gallos et al. (2007)

Quantify anticorrelation between hubsat all length scales

hubs

hubs

Renormalize

Hubs connected directly

Hub-Hub Correlation function: fraction of hub-hub connections

Hub-hub correlations organized in a self-similar way

The larger de implies more anticorrelations

(fractal) (non-fractal)

Anticorrelations are essential for fractal structures

non-fractal

fractal

Exponent de determines the joint probability distribution

What is the origin of fractality?

• very compact networks• hubs connected with other hubs• strong hub-hub “attraction”• assortativity

Non-fractal networks

• less compact networks• hubs connected with non-hubs• strong hub-hub “repulsion”• dissasortativity

Fractal networks

Internet, socialAll available models: BA model, hierarchicalrandom scale free, JKK, etc

WWW, PIN, metabolic, genetic, neural networks, protein homology, taxonomy

How to model it? renormalization reverses time evolution

Mode IIMode I

tim

e

Both mass and degree increase exponentially with time

Scale-free:

offspring nodes attached to their parents

(m=2) in this case

reno

rmal

ize

Song, Havlin, Makse, Nature Physics, 2006

How does the length increase with time?

Mode II: FRACTALMode I: NONFRACTAL

SMALL WORLD

Combine two modes together

tim

e

e=0.5

Mode I with probability e Mode II with probability 1-e

reno

rmal

ize

m = 2

Model

A multiplicative growth processof the number of nodes and links

Probability ehubs always connected

strong hub attractionshould lead to non-fractal

Probability 1-ehubs never connectedstrong hub repulsionshould lead to fractal

Analogous to duplication/divergence

mechanism in proteins??

For the both models, each step the total number of nodes scale as n = 2m +1( N(t+1) = nN(t) ). Now we investigate the transformation of the lengths. They show quite different ways for this two models as following:

Then we lead to two different scaling law of N ~ L

Mode III: L(t+1) =3L(t)Mode II: L(t+1) = 2L(t)+1

Mode I: L(t+1) = L(t)+2

smaller

smaller

Different growth modes lead to differenttopologies

Suppose we have e probability to have mode I, 1-e probability to have mode II and mode III. Then we have:

or

Dynamical model

Model predicts all exponents in terms ofgrowth rates

Each step the total mass scales with a constant n, all the degrees scale with a constant s.

The length scales with a constant a, we obtain:

We predict the fractal exponents:

PredictionsModel reproduces local small world, scale-free and fractality

NON-FRACTAL• attraction between hubs• non-fractal• small world globally

FRACTAL• repulsion between hubs leads to fractal topology• small world locally inside well defined communities

yeast

The model reproduces the main features of real networks

Case 1: e = 0.8: FRACTALS Case 2: e = 1.0: NON-FRACTALS

Summary of scaling exponents and scaling relationsMass:

Links:

Hub-hub correlations:

Modularity ratioModularity exponent:

Number of hub-hub links

Number of links outside modules

Number of links inside modules

Modularity is also scale-invariant

Protein Homology

Similarities between sequence ofamino-acids (BLAST)Network of 5 million proteins1.2 TB of data growing at 50GBper month. Adai et al. J Mol Biol (2004)

Yeast protein interaction

Large modularity Ultramodularity

Time evolution in yeast network

Multiplicative and exponential growth in yeast PINLength-scales, number of conserved proteins and degree

Self-similar learning dynamics of the brainCalcium imaging of spontaneous action potentials in large neuronal populations of a slice of the medial prefrontal cortex of a brain slice of mouse.

QuickTime™ and aSorenson Video 3 decompressorare needed to see this picture.

Rafael Yuste and IkegayaJohn Cageminimalistavant-gardemusic

t = 15 sec t = 30 sec t = 45 sec t = 60 sec

t = 75 sec t = 90 sec t = 105 sec t = 120 sec

Time evolution of the network

The degree distribution P(k) is invariant under evolution. The plots go from 30 sec to 120 sec

The fractal dimension is also invariant under evolution from 30 sec to 120 sec

The degree exponents and fractal dimension are invariant under the time evolution

Scale-transformation of degree

We verify the formula:

k(t1) = S(t1|t2) k(t2)

Here we fix t2 = 120 sec, and take t1 from 30 sec to 105 sec. The linear dependency is verified for different times t1.

From the theory: N(t) = s(t)γ−

The inset shows that both N(t) and s(t) increase exponentially:

N(t) ~ exp(0.014t)

s(t) ~ exp(0.021t)

This gives rise to the following scaling relation:

Confirmation of the scaling formula for the degree exponent as a function of the fractal exponents

Tolerance of the network under random failureand intentional attack

We plot the largest cluster size as a function of the fraction p of nodes removed

A new principle of network dynamics 1930solid-state physicsbig world

1960Erdos-Renyi model small world

democracy=socialism

1999BA model “rich-get-richer”=

capitalism

2005fractal model of modularity

“rich-get-richer” at the expense of the “poor”=

globalizationLess vulnerable to intentional attacks:Designed by Evolutionary pressure.

Summary

• In contrast to common belief, many real world networks are self-similar.• FRACTALS: WWW, Protein interactions, metabolic networks, neural networks, homology networks, tree of life. • NON-FRACTALS: Internet, social, all models.• Communities/modules are self-similar, as well.• Scaling theory describes the dynamical evolution.• Boxes are related to the functional modules in metabolic and protein networks.• Origin of self similarity: anticorrelation between hubs• Fractal networks are less vulnerable than non-fractal networks

Graph theoretical representation of a metabolicGraph theoretical representation of a metabolicnetworknetwork

(a) A (a) A pathway (catalyzed by Mg2+-dependant enzymes).(b) All interacting metabolites are considered equally. (c) For many biological applications it is useful to ignore co-factors, such as the high energy-phosphate donor ATP, which results in a second type of mapping that connects only the main source metabolites to the main products.

More topological fractals

WWW

nd.edu domain

Hollywood film actors

212,000 actors

300,000 web-pages

Burning algorithms

Compact box burning: CBBSong et al. JSTAT (2007)

Maximum excluded mass burning: MEMBBurning from the hubs with the radius r

Minimazing the number of boxes is analogous to maximizing the mass of each box: Modularity